Per this comment, it looks like we don’t want to downgrade pytorch like that. Anyone know how or if we should upgrade given this issue? I’m not so keen to go from a cuda8.0-based pytorch to a cuda7.5-based one.
I’m also not sure how I initially got a cuda8.0-based pytorch. I’m pretty sure I installed using conda with the environment.yml file. ¯\_(ツ)_/¯
conda update --all is going to look at your environment.yml for instructions on what to update. If you aren’t specifying the specific version or wildcarding major/minor versions alongside of the package, it’s going to want to update them all to the latest version.
So I’m kinda conflicted. I can see where it could mess things up, but also where it could fix things in the case of minor improvements (e.g. bug fixes, enhancements, etc…) that aren’t breaking changes.